registry.py 47.9 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
5
6
"""
Whenever you add an architecture to this page, please also update
`tests/models/registry.py` with example HuggingFace models for it.
"""
7

8
import importlib
9
import json
10
import os
11
import pickle
12
13
import subprocess
import sys
14
import tempfile
15
from abc import ABC, abstractmethod
16
from collections.abc import Callable, Set
17
from dataclasses import asdict, dataclass, field
18
from functools import lru_cache
19
from pathlib import Path
20
from typing import TYPE_CHECKING, Any, TypeVar
21
22

import torch.nn as nn
23
import transformers
24

25
from vllm import envs
26
27
28
29
30
from vllm.config import (
    ModelConfig,
    iter_architecture_defaults,
    try_match_architecture_defaults,
)
31
from vllm.logger import init_logger
32
from vllm.logging_utils import logtime
33
from vllm.transformers_utils.dynamic_module import try_get_class_from_dynamic_module
34
from vllm.utils.hashing import safe_hash
35

36
37
if TYPE_CHECKING:
    from vllm.config.model import AttnTypeStr
38
    from vllm.config.pooler import SequencePoolingType, TokenPoolingType
39
40
else:
    AttnTypeStr = Any
41
42
    SequencePoolingType = Any
    TokenPoolingType = Any
43
44


45
46
47
48
49
from .interfaces import (
    has_inner_state,
    has_noops,
    is_attention_free,
    is_hybrid,
Patrick von Platen's avatar
Patrick von Platen committed
50
    requires_raw_input_tokens,
51
    supports_cross_encoding,
52
    supports_mamba_prefix_caching,
53
54
55
56
57
58
59
    supports_multimodal,
    supports_multimodal_encoder_tp_data,
    supports_multimodal_raw_input_only,
    supports_pp,
    supports_transcription,
)
from .interfaces_base import (
60
    get_attn_type,
61
62
    get_default_seq_pooling_type,
    get_default_tok_pooling_type,
63
64
65
    is_pooling_model,
    is_text_generation_model,
)
66
67
68

logger = init_logger(__name__)

69
70
_TEXT_GENERATION_MODELS = {
    # [Decoder-only]
71
    "AfmoeForCausalLM": ("afmoe", "AfmoeForCausalLM"),
72
    "ApertusForCausalLM": ("apertus", "ApertusForCausalLM"),
73
74
    "AquilaModel": ("llama", "LlamaForCausalLM"),
    "AquilaForCausalLM": ("llama", "LlamaForCausalLM"),  # AquilaChat2
Raghav Ravishankar's avatar
Raghav Ravishankar committed
75
    "ArceeForCausalLM": ("arcee", "ArceeForCausalLM"),
76
    "ArcticForCausalLM": ("arctic", "ArcticForCausalLM"),
77
78
79
80
    # baichuan-7b, upper case 'C' in the class name
    "BaiChuanForCausalLM": ("baichuan", "BaiChuanForCausalLM"),
    # baichuan-13b, lower case 'c' in the class name
    "BaichuanForCausalLM": ("baichuan", "BaichuanForCausalLM"),
81
    "BailingMoeForCausalLM": ("bailing_moe", "BailingMoeForCausalLM"),
ant-yy's avatar
ant-yy committed
82
    "BailingMoeV2ForCausalLM": ("bailing_moe", "BailingMoeV2ForCausalLM"),
Yu Chin Fabian Lim's avatar
Yu Chin Fabian Lim committed
83
    "BambaForCausalLM": ("bamba", "BambaForCausalLM"),
84
    "BloomForCausalLM": ("bloom", "BloomForCausalLM"),
85
    "ChatGLMModel": ("chatglm", "ChatGLMForCausalLM"),
86
    "ChatGLMForConditionalGeneration": ("chatglm", "ChatGLMForCausalLM"),
87
    "CohereForCausalLM": ("commandr", "CohereForCausalLM"),
88
    "Cohere2ForCausalLM": ("commandr", "CohereForCausalLM"),
89
    "CwmForCausalLM": ("llama", "LlamaForCausalLM"),
90
    "DbrxForCausalLM": ("dbrx", "DbrxForCausalLM"),
91
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
92
    "DeepseekForCausalLM": ("deepseek_v2", "DeepseekForCausalLM"),
93
    "DeepseekV2ForCausalLM": ("deepseek_v2", "DeepseekV2ForCausalLM"),
94
    "DeepseekV3ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
95
    "DeepseekV32ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
96
    "Dots1ForCausalLM": ("dots1", "Dots1ForCausalLM"),
97
    "Ernie4_5ForCausalLM": ("ernie45", "Ernie4_5ForCausalLM"),
98
    "Ernie4_5_MoeForCausalLM": ("ernie45_moe", "Ernie4_5_MoeForCausalLM"),
99
    "ExaoneForCausalLM": ("exaone", "ExaoneForCausalLM"),
100
    "Exaone4ForCausalLM": ("exaone4", "Exaone4ForCausalLM"),
Kyungmin Lee's avatar
Kyungmin Lee committed
101
    "ExaoneMoEForCausalLM": ("exaone_moe", "ExaoneMoeForCausalLM"),
102
    "Fairseq2LlamaForCausalLM": ("fairseq2_llama", "Fairseq2LlamaForCausalLM"),
103
104
105
    "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
    "FalconMambaForCausalLM": ("mamba", "MambaForCausalLM"),
    "FalconH1ForCausalLM": ("falcon_h1", "FalconH1ForCausalLM"),
106
    "FlexOlmoForCausalLM": ("flex_olmo", "FlexOlmoForCausalLM"),
107
108
    "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),
    "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
109
    "Gemma3ForCausalLM": ("gemma3", "Gemma3ForCausalLM"),
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
110
    "Gemma3nForCausalLM": ("gemma3n", "Gemma3nForCausalLM"),
111
    "Qwen3NextForCausalLM": ("qwen3_next", "Qwen3NextForCausalLM"),
112
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
113
    "Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
114
    "Glm4MoeForCausalLM": ("glm4_moe", "Glm4MoeForCausalLM"),
115
    "Glm4MoeLiteForCausalLM": ("glm4_moe_lite", "Glm4MoeLiteForCausalLM"),
116
    "GptOssForCausalLM": ("gpt_oss", "GptOssForCausalLM"),
117
118
119
120
121
122
    "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
    "GPTBigCodeForCausalLM": ("gpt_bigcode", "GPTBigCodeForCausalLM"),
    "GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),
    "GPTNeoXForCausalLM": ("gpt_neox", "GPTNeoXForCausalLM"),
    "GraniteForCausalLM": ("granite", "GraniteForCausalLM"),
    "GraniteMoeForCausalLM": ("granitemoe", "GraniteMoeForCausalLM"),
123
124
    "GraniteMoeHybridForCausalLM": ("granitemoehybrid", "GraniteMoeHybridForCausalLM"),  # noqa: E501
    "GraniteMoeSharedForCausalLM": ("granitemoeshared", "GraniteMoeSharedForCausalLM"),  # noqa: E501
125
    "GritLM": ("gritlm", "GritLM"),
Bijaya Dangol's avatar
Bijaya Dangol committed
126
127
    "Grok1ModelForCausalLM": ("grok1", "GrokForCausalLM"),
    "Grok1ForCausalLM": ("grok1", "GrokForCausalLM"),
128
129
    "HunYuanMoEV1ForCausalLM": ("hunyuan_v1", "HunYuanMoEV1ForCausalLM"),
    "HunYuanDenseV1ForCausalLM": ("hunyuan_v1", "HunYuanDenseV1ForCausalLM"),
130
    "HCXVisionForCausalLM": ("hyperclovax_vision", "HCXVisionForCausalLM"),
131
132
    "InternLMForCausalLM": ("llama", "LlamaForCausalLM"),
    "InternLM2ForCausalLM": ("internlm2", "InternLM2ForCausalLM"),
133
    "InternLM2VEForCausalLM": ("internlm2_ve", "InternLM2VEForCausalLM"),
134
    "InternLM3ForCausalLM": ("llama", "LlamaForCausalLM"),
135
136
    "IQuestCoderForCausalLM": ("llama", "LlamaForCausalLM"),
    "IQuestLoopCoderForCausalLM": ("iquest_loopcoder", "IQuestLoopCoderForCausalLM"),
137
    "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
138
    "Jais2ForCausalLM": ("jais2", "Jais2ForCausalLM"),
139
    "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
140
    "KimiLinearForCausalLM": ("kimi_linear", "KimiLinearForCausalLM"),  # noqa: E501
141
    "Lfm2ForCausalLM": ("lfm2", "Lfm2ForCausalLM"),
Paul Pak's avatar
Paul Pak committed
142
    "Lfm2MoeForCausalLM": ("lfm2_moe", "Lfm2MoeForCausalLM"),
143
    "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
144
    "Llama4ForCausalLM": ("llama4", "Llama4ForCausalLM"),
145
146
    # For decapoda-research/llama-*
    "LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
XuruiYang's avatar
XuruiYang committed
147
    "LongcatFlashForCausalLM": ("longcat_flash", "LongcatFlashForCausalLM"),
148
    "MambaForCausalLM": ("mamba", "MambaForCausalLM"),
149
    "Mamba2ForCausalLM": ("mamba2", "Mamba2ForCausalLM"),
150
151
    "MiniCPMForCausalLM": ("minicpm", "MiniCPMForCausalLM"),
    "MiniCPM3ForCausalLM": ("minicpm3", "MiniCPM3ForCausalLM"),
152
153
154
    "MiniMaxForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxText01ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxM1ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
155
    "MiniMaxM2ForCausalLM": ("minimax_m2", "MiniMaxM2ForCausalLM"),
156
    "MistralForCausalLM": ("mistral", "MistralForCausalLM"),
157
    "MistralLarge3ForCausalLM": ("mistral_large_3", "MistralLarge3ForCausalLM"),
158
159
160
161
    "MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),
    # transformers's mpt class has lower case
    "MptForCausalLM": ("mpt", "MPTForCausalLM"),
    "MPTForCausalLM": ("mpt", "MPTForCausalLM"),
162
    "MiMoForCausalLM": ("mimo", "MiMoForCausalLM"),
163
    "MiMoV2FlashForCausalLM": ("mimo_v2_flash", "MiMoV2FlashForCausalLM"),
164
    "NemotronForCausalLM": ("nemotron", "NemotronForCausalLM"),
Luis Vega's avatar
Luis Vega committed
165
    "NemotronHForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
166
    "OlmoForCausalLM": ("olmo", "OlmoForCausalLM"),
167
    "Olmo2ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
168
    "Olmo3ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
169
170
171
    "OlmoeForCausalLM": ("olmoe", "OlmoeForCausalLM"),
    "OPTForCausalLM": ("opt", "OPTForCausalLM"),
    "OrionForCausalLM": ("orion", "OrionForCausalLM"),
172
    "OuroForCausalLM": ("ouro", "OuroForCausalLM"),
173
    "PanguEmbeddedForCausalLM": ("openpangu", "PanguEmbeddedForCausalLM"),
174
    "PanguProMoEV2ForCausalLM": ("openpangu", "PanguProMoEV2ForCausalLM"),
175
    "PanguUltraMoEForCausalLM": ("openpangu", "PanguUltraMoEForCausalLM"),
176
177
178
179
    "PersimmonForCausalLM": ("persimmon", "PersimmonForCausalLM"),
    "PhiForCausalLM": ("phi", "PhiForCausalLM"),
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
    "PhiMoEForCausalLM": ("phimoe", "PhiMoEForCausalLM"),
Shinichi Hemmi's avatar
Shinichi Hemmi committed
180
    "Plamo2ForCausalLM": ("plamo2", "Plamo2ForCausalLM"),
181
    "Plamo3ForCausalLM": ("plamo3", "Plamo3ForCausalLM"),
182
    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
183
184
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
185
186
    "Qwen3ForCausalLM": ("qwen3", "Qwen3ForCausalLM"),
    "Qwen3MoeForCausalLM": ("qwen3_moe", "Qwen3MoeForCausalLM"),
187
    "RWForCausalLM": ("falcon", "FalconForCausalLM"),
188
    "SeedOssForCausalLM": ("seed_oss", "SeedOssForCausalLM"),
Li Xie's avatar
Li Xie committed
189
    "Step1ForCausalLM": ("step1", "Step1ForCausalLM"),
Song's avatar
Song committed
190
    "Step3TextForCausalLM": ("step3_text", "Step3TextForCausalLM"),
191
192
193
194
    "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
    "SolarForCausalLM": ("solar", "SolarForCausalLM"),
195
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
196
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
197
    "TeleFLMForCausalLM": ("teleflm", "TeleFLMForCausalLM"),
198
    "XverseForCausalLM": ("llama", "LlamaForCausalLM"),
199
    "Zamba2ForCausalLM": ("zamba2", "Zamba2ForCausalLM"),
200
201
202
}

_EMBEDDING_MODELS = {
203
    # [Text-only]
204
    "BertModel": ("bert", "BertEmbeddingModel"),
205
    "BertSpladeSparseEmbeddingModel": ("bert", "BertSpladeSparseEmbeddingModel"),
206
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
207
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
208
    "Gemma3TextModel": ("gemma3", "Gemma3Model"),
209
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
210
    "GPT2ForSequenceClassification": ("gpt2", "GPT2ForSequenceClassification"),
211
    "GritLM": ("gritlm", "GritLM"),
212
213
    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
    "GteNewModel": ("bert_with_rope", "GteNewModel"),
214
    "InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
215
    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),  # noqa: E501
216
    "LlamaBidirectionalModel": ("llama", "LlamaBidirectionalModel"),
217
    "LlamaModel": ("llama", "LlamaForCausalLM"),
218
219
    **{
        # Multiple models share the same architecture, so we include them all
220
221
        k: (mod, arch)
        for k, (mod, arch) in _TEXT_GENERATION_MODELS.items()
222
223
        if arch == "LlamaForCausalLM"
    },
224
    "MistralModel": ("llama", "LlamaForCausalLM"),
225
    "ModernBertModel": ("modernbert", "ModernBertModel"),
226
    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
227
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
228
    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
229
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
230
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
231
    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
232
233
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
234
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
235
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
236
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
237
    "BgeM3EmbeddingModel": ("roberta", "BgeM3EmbeddingModel"),
238
    # [Multimodal]
239
    "CLIPModel": ("clip", "CLIPEmbeddingModel"),
240
241
242
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
243
    ),
244
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
245
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
246
    "SiglipModel": ("siglip", "SiglipEmbeddingModel"),
247
248
    # Technically Terratorch models work on images, both in
    # input and output. I am adding it here because it piggy-backs on embedding
249
    # models for the time being.
250
251
    "PrithviGeoSpatialMAE": ("terratorch", "Terratorch"),
    "Terratorch": ("terratorch", "Terratorch"),
252
253
}

254
255
_CROSS_ENCODER_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
256
    "BertForTokenClassification": ("bert", "BertForTokenClassification"),
257
258
259
260
    "GteNewForSequenceClassification": (
        "bert_with_rope",
        "GteNewForSequenceClassification",
    ),
261
262
263
264
265
    "JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"),
    "LlamaBidirectionalForSequenceClassification": (
        "llama",
        "LlamaBidirectionalForSequenceClassification",
    ),
266
267
268
269
    "ModernBertForSequenceClassification": (
        "modernbert",
        "ModernBertForSequenceClassification",
    ),
270
271
272
273
    "ModernBertForTokenClassification": (
        "modernbert",
        "ModernBertForTokenClassification",
    ),
274
275
276
277
278
    "RobertaForSequenceClassification": ("roberta", "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": (
        "roberta",
        "RobertaForSequenceClassification",
    ),
279
280
}

281
_MULTIMODAL_MODELS = {
282
    # [Decoder-only]
283
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
284
285
286
287
    "AudioFlamingo3ForConditionalGeneration": (
        "audioflamingo3",
        "AudioFlamingo3ForConditionalGeneration",
    ),
288
289
290
    "AyaVisionForConditionalGeneration": (
        "aya_vision",
        "AyaVisionForConditionalGeneration",
291
    ),
292
    "BagelForConditionalGeneration": ("bagel", "BagelForConditionalGeneration"),
293
    "BeeForConditionalGeneration": ("bee", "BeeForConditionalGeneration"),
294
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
295
296
297
    "ChameleonForConditionalGeneration": (
        "chameleon",
        "ChameleonForConditionalGeneration",
298
    ),
299
300
301
    "Cohere2VisionForConditionalGeneration": (
        "cohere2_vision",
        "Cohere2VisionForConditionalGeneration",
302
    ),
303
    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
304
    "DeepseekOCRForCausalLM": ("deepseek_ocr", "DeepseekOCRForCausalLM"),
Roger Wang's avatar
Roger Wang committed
305
    "DotsOCRForCausalLM": ("dots_ocr", "DotsOCRForCausalLM"),
306
307
308
309
    "Eagle2_5_VLForConditionalGeneration": (
        "eagle2_5_vl",
        "Eagle2_5_VLForConditionalGeneration",
    ),
310
311
312
    "Ernie4_5_VLMoeForConditionalGeneration": (
        "ernie45_vl",
        "Ernie4_5_VLMoeForConditionalGeneration",
313
    ),
314
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
315
    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501
316
317
318
    "Gemma3nForConditionalGeneration": (
        "gemma3n_mm",
        "Gemma3nForConditionalGeneration",
319
    ),
320
    "GlmAsrForConditionalGeneration": ("glmasr", "GlmAsrForConditionalGeneration"),
321
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
322
    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),  # noqa: E501
Jee Jee Li's avatar
Jee Jee Li committed
323
    "Glm4vMoeForConditionalGeneration": ("glm4_1v", "Glm4vMoeForConditionalGeneration"),  # noqa: E501
324
325
326
    "GraniteSpeechForConditionalGeneration": (
        "granite_speech",
        "GraniteSpeechForConditionalGeneration",
327
    ),
328
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
329
330
331
332
    "HunYuanVLForConditionalGeneration": (
        "hunyuan_vision",
        "HunYuanVLForConditionalGeneration",
    ),
ltd0924's avatar
ltd0924 committed
333
    "StepVLForConditionalGeneration": ("step_vl", "StepVLForConditionalGeneration"),
334
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
335
    "NemotronH_Nano_VL_V2": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
Zero's avatar
Zero committed
336
337
338
339
    "OpenCUAForConditionalGeneration": (
        "opencua",
        "OpenCUAForConditionalGeneration",
    ),
340
341
342
    "InternS1ForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
343
    ),
344
345
346
    "InternVLForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
347
    ),
348
349
350
351
    "Idefics3ForConditionalGeneration": (
        "idefics3",
        "Idefics3ForConditionalGeneration",
    ),
oscardev256's avatar
oscardev256 committed
352
    "IsaacForConditionalGeneration": ("isaac", "IsaacForConditionalGeneration"),
353
    "SmolVLMForConditionalGeneration": ("smolvlm", "SmolVLMForConditionalGeneration"),  # noqa: E501
354
    "KananaVForConditionalGeneration": ("kanana_v", "KananaVForConditionalGeneration"),
355
    "KeyeForConditionalGeneration": ("keye", "KeyeForConditionalGeneration"),
356
357
358
    "KeyeVL1_5ForConditionalGeneration": (
        "keye_vl1_5",
        "KeyeVL1_5ForConditionalGeneration",
359
    ),
360
    "RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
361
    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),  # noqa: E501
Roger Wang's avatar
Roger Wang committed
362
    "KimiK25ForConditionalGeneration": ("kimi_k25", "KimiK25ForConditionalGeneration"),  # noqa: E501
363
364
365
366
    "LightOnOCRForConditionalGeneration": (
        "lightonocr",
        "LightOnOCRForConditionalGeneration",
    ),
367
    "Lfm2VlForConditionalGeneration": ("lfm2_vl", "Lfm2VLForConditionalGeneration"),
368
    "Llama_Nemotron_Nano_VL": ("nemotron_vl", "LlamaNemotronVLChatModel"),
369
    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),  # noqa: E501
370
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
371
372
373
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
374
    ),
375
376
377
    "LlavaNextVideoForConditionalGeneration": (
        "llava_next_video",
        "LlavaNextVideoForConditionalGeneration",
378
    ),
379
380
381
    "LlavaOnevisionForConditionalGeneration": (
        "llava_onevision",
        "LlavaOnevisionForConditionalGeneration",
382
    ),
383
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),  # noqa: E501
384
    "MiDashengLMModel": ("midashenglm", "MiDashengLMModel"),
385
386
387
    "MiniMaxVL01ForConditionalGeneration": (
        "minimax_vl_01",
        "MiniMaxVL01ForConditionalGeneration",
388
    ),
389
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
390
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
391
392
393
    "Mistral3ForConditionalGeneration": (
        "mistral3",
        "Mistral3ForConditionalGeneration",
394
    ),
395
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
396
    "Molmo2ForConditionalGeneration": ("molmo2", "Molmo2ForConditionalGeneration"),
397
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
398
    "Ovis": ("ovis", "Ovis"),
399
    "Ovis2_5": ("ovis2_5", "Ovis2_5"),
400
401
402
403
    "PaddleOCRVLForConditionalGeneration": (
        "paddleocr_vl",
        "PaddleOCRVLForConditionalGeneration",
    ),
404
405
406
407
    "PaliGemmaForConditionalGeneration": (
        "paligemma",
        "PaliGemmaForConditionalGeneration",
    ),
408
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
409
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
410
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
411
    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),  # noqa: E501
412
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
413
414
415
    "Qwen2_5_VLForConditionalGeneration": (
        "qwen2_5_vl",
        "Qwen2_5_VLForConditionalGeneration",
416
    ),
417
418
419
    "Qwen2AudioForConditionalGeneration": (
        "qwen2_audio",
        "Qwen2AudioForConditionalGeneration",
420
    ),
421
422
423
    "Qwen2_5OmniModel": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
424
    ),
425
426
427
    "Qwen2_5OmniForConditionalGeneration": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
428
    ),
429
430
431
432
    "Qwen3OmniMoeForConditionalGeneration": (
        "qwen3_omni_moe_thinker",
        "Qwen3OmniMoeThinkerForConditionalGeneration",
    ),
433
    "Qwen3VLForConditionalGeneration": ("qwen3_vl", "Qwen3VLForConditionalGeneration"),  # noqa: E501
434
435
436
    "Qwen3VLMoeForConditionalGeneration": (
        "qwen3_vl_moe",
        "Qwen3VLMoeForConditionalGeneration",
437
    ),
438
    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
Song's avatar
Song committed
439
    "Step3VLForConditionalGeneration": ("step3_vl", "Step3VLForConditionalGeneration"),  # noqa: E501
汪志鹏's avatar
汪志鹏 committed
440
    "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),  # noqa: E501
441
442
443
    "Tarsier2ForConditionalGeneration": (
        "qwen2_vl",
        "Tarsier2ForConditionalGeneration",
444
    ),
445
    "UltravoxModel": ("ultravox", "UltravoxModel"),
Patrick von Platen's avatar
Patrick von Platen committed
446
    "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),  # noqa: E501
Patrick von Platen's avatar
Patrick von Platen committed
447
    "VoxtralStreamingGeneration": ("voxtral_streaming", "VoxtralStreamingGeneration"),  # noqa: E501
448
    # [Encoder-decoder]
449
450
451
452
    "NemotronParseForConditionalGeneration": (
        "nemotron_parse",
        "NemotronParseForConditionalGeneration",
    ),
453
    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),  # noqa: E501
454
}
455
456

_SPECULATIVE_DECODING_MODELS = {
457
    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
458
    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
zhiweiz's avatar
zhiweiz committed
459
    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
460
    "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
461
    "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
462
    "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
463
    "Eagle3Qwen2_5vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
464
    "Eagle3Qwen3vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
465
466
467
468
    "EagleMistralLarge3ForCausalLM": (
        "mistral_large_3_eagle",
        "EagleMistralLarge3ForCausalLM",
    ),
469
    "EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
470
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
471
    "ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
Kyungmin Lee's avatar
Kyungmin Lee committed
472
    "ExaoneMoeMTP": ("exaone_moe_mtp", "ExaoneMoeMTP"),
XuruiYang's avatar
XuruiYang committed
473
    "LongCatFlashMTPModel": ("longcat_flash_mtp", "LongCatFlashMTP"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
474
    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
475
    "Glm4MoeLiteMTPModel": ("glm4_moe_lite_mtp", "Glm4MoeLiteMTP"),
476
    "MedusaModel": ("medusa", "Medusa"),
477
    "OpenPanguMTPModel": ("openpangu_mtp", "OpenPanguMTP"),
478
    "Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
479
480
481
    # Temporarily disabled.
    # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
    # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
482
}
483

484
_TRANSFORMERS_SUPPORTED_MODELS = {
485
486
487
    # Text generation models
    "SmolLM3ForCausalLM": ("transformers", "TransformersForCausalLM"),
    # Multimodal models
488
489
490
491
    "Emu3ForConditionalGeneration": (
        "transformers",
        "TransformersMultiModalForCausalLM",
    ),
492
493
494
}

_TRANSFORMERS_BACKEND_MODELS = {
495
    # Text generation models
496
    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
    "TransformersMoEForCausalLM": ("transformers", "TransformersMoEForCausalLM"),
    # Multimodal models
    "TransformersMultiModalForCausalLM": (
        "transformers",
        "TransformersMultiModalForCausalLM",
    ),
    "TransformersMultiModalMoEForCausalLM": (
        "transformers",
        "TransformersMultiModalMoEForCausalLM",
    ),
    # Embedding models
    "TransformersEmbeddingModel": ("transformers", "TransformersEmbeddingModel"),
    "TransformersMoEEmbeddingModel": ("transformers", "TransformersMoEEmbeddingModel"),
    "TransformersMultiModalEmbeddingModel": (
        "transformers",
        "TransformersMultiModalEmbeddingModel",
    ),
    # Sequence classification models
515
    "TransformersForSequenceClassification": (
516
        "transformers",
517
        "TransformersForSequenceClassification",
518
    ),
519
    "TransformersMoEForSequenceClassification": (
520
        "transformers",
521
        "TransformersMoEForSequenceClassification",
522
    ),
523
524
525
    "TransformersMultiModalForSequenceClassification": (
        "transformers",
        "TransformersMultiModalForSequenceClassification",
526
    ),
527
}
528

529
_VLLM_MODELS = {
530
    **_TEXT_GENERATION_MODELS,
531
    **_EMBEDDING_MODELS,
532
    **_CROSS_ENCODER_MODELS,
533
    **_MULTIMODAL_MODELS,
534
    **_SPECULATIVE_DECODING_MODELS,
535
536
    **_TRANSFORMERS_SUPPORTED_MODELS,
    **_TRANSFORMERS_BACKEND_MODELS,
537
538
}

539
540
541
542
# This variable is used as the args for subprocess.run(). We
# can modify  this variable to alter the args if needed. e.g.
# when we use par format to pack things together, sys.executable
# might not be the target we want to run.
543
_SUBPROCESS_COMMAND = [sys.executable, "-m", "vllm.model_executor.models.registry"]
544

545
_PREVIOUSLY_SUPPORTED_MODELS = {
546
    "MotifForCausalLM": "0.10.2",
547
    "Phi3SmallForCausalLM": "0.9.2",
548
    "Phi4FlashForCausalLM": "0.10.2",
549
    "Phi4MultimodalForCausalLM": "0.12.0",
550
551
552
553
554
555
556
557
558
    # encoder-decoder models except whisper
    # have been removed for V0 deprecation.
    "BartModel": "0.10.2",
    "BartForConditionalGeneration": "0.10.2",
    "DonutForConditionalGeneration": "0.10.2",
    "Florence2ForConditionalGeneration": "0.10.2",
    "MBartForConditionalGeneration": "0.10.2",
    "MllamaForConditionalGeneration": "0.10.2",
}
559

560

561
562
@dataclass(frozen=True)
class _ModelInfo:
563
    architecture: str
564
    is_text_generation_model: bool
565
    is_pooling_model: bool
566
    attn_type: AttnTypeStr
567
568
    default_seq_pooling_type: SequencePoolingType
    default_tok_pooling_type: TokenPoolingType
569
    supports_cross_encoding: bool
570
    supports_multimodal: bool
571
    supports_multimodal_raw_input_only: bool
Patrick von Platen's avatar
Patrick von Platen committed
572
    requires_raw_input_tokens: bool
573
    supports_multimodal_encoder_tp_data: bool
574
    supports_pp: bool
575
576
    has_inner_state: bool
    is_attention_free: bool
577
    is_hybrid: bool
578
    has_noops: bool
579
    supports_mamba_prefix_caching: bool
580
    supports_transcription: bool
581
    supports_transcription_only: bool
582
583

    @staticmethod
584
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
585
        return _ModelInfo(
586
            architecture=model.__name__,
587
            is_text_generation_model=is_text_generation_model(model),
588
            is_pooling_model=is_pooling_model(model),
589
590
            default_seq_pooling_type=get_default_seq_pooling_type(model),
            default_tok_pooling_type=get_default_tok_pooling_type(model),
591
            attn_type=get_attn_type(model),
592
            supports_cross_encoding=supports_cross_encoding(model),
593
            supports_multimodal=supports_multimodal(model),
594
595
596
            supports_multimodal_raw_input_only=supports_multimodal_raw_input_only(
                model
            ),
Patrick von Platen's avatar
Patrick von Platen committed
597
            requires_raw_input_tokens=requires_raw_input_tokens(model),
598
599
600
            supports_multimodal_encoder_tp_data=supports_multimodal_encoder_tp_data(
                model
            ),
601
            supports_pp=supports_pp(model),
602
603
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
604
            is_hybrid=is_hybrid(model),
605
            supports_mamba_prefix_caching=supports_mamba_prefix_caching(model),
606
            supports_transcription=supports_transcription(model),
607
608
609
            supports_transcription_only=(
                supports_transcription(model) and model.supports_transcription_only
            ),
610
            has_noops=has_noops(model),
611
        )
612
613


614
615
616
617
class _BaseRegisteredModel(ABC):
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
618

619
    @abstractmethod
620
    def load_model_cls(self) -> type[nn.Module]:
621
        raise NotImplementedError
622
623


624
625
626
627
628
629
630
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
631
    model_cls: type[nn.Module]
632
633

    @staticmethod
634
    def from_model_cls(model_cls: type[nn.Module]):
635
636
637
638
639
640
641
642
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

    def inspect_model_cls(self) -> _ModelInfo:
        return self.interfaces

643
    def load_model_cls(self) -> type[nn.Module]:
644
645
646
647
648
649
650
651
        return self.model_cls


@dataclass(frozen=True)
class _LazyRegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has not been imported in the main process.
    """
652

653
654
655
    module_name: str
    class_name: str

656
657
658
659
660
661
662
663
    @staticmethod
    def _get_cache_dir() -> Path:
        return Path(envs.VLLM_CACHE_ROOT) / "modelinfos"

    def _get_cache_filename(self) -> str:
        cls_name = f"{self.module_name}-{self.class_name}".replace(".", "-")
        return f"{cls_name}.json"

664
    def _load_modelinfo_from_cache(self, module_hash: str) -> _ModelInfo | None:
665
666
        try:
            try:
667
                modelinfo_path = self._get_cache_dir() / self._get_cache_filename()
668
669
670
                with open(modelinfo_path, encoding="utf-8") as file:
                    mi_dict = json.load(file)
            except FileNotFoundError:
671
                logger.debug(
672
                    "Cached model info file for class %s.%s not found",
673
674
675
                    self.module_name,
                    self.class_name,
                )
676
677
678
                return None

            if mi_dict["hash"] != module_hash:
679
                logger.debug(
680
                    "Cached model info file for class %s.%s is stale",
681
682
683
                    self.module_name,
                    self.class_name,
                )
684
685
686
687
688
                return None

            # file not changed, use cached _ModelInfo properties
            return _ModelInfo(**mi_dict["modelinfo"])
        except Exception:
689
            logger.debug(
690
                "Cached model info for class %s.%s error. ",
691
692
693
                self.module_name,
                self.class_name,
            )
694
695
            return None

696
    def _save_modelinfo_to_cache(self, mi: _ModelInfo, module_hash: str) -> None:
697
698
        """save dictionary json file to cache"""
        from vllm.model_executor.model_loader.weight_utils import atomic_writer
699

700
701
702
703
704
705
706
707
        try:
            modelinfo_dict = {
                "hash": module_hash,
                "modelinfo": asdict(mi),
            }
            cache_dir = self._get_cache_dir()
            cache_dir.mkdir(parents=True, exist_ok=True)
            modelinfo_path = cache_dir / self._get_cache_filename()
708
            with atomic_writer(modelinfo_path, encoding="utf-8") as f:
709
710
711
712
713
                json.dump(modelinfo_dict, f, indent=2)
        except Exception:
            logger.exception("Error saving model info cache.")

    @logtime(logger=logger, msg="Registry inspect model class")
714
    def inspect_model_cls(self) -> _ModelInfo:
715
        model_path = Path(__file__).parent / f"{self.module_name.split('.')[-1]}.py"
716
        module_hash = None
717

718
719
        if model_path.exists():
            with open(model_path, "rb") as f:
720
                module_hash = safe_hash(f.read(), usedforsecurity=False).hexdigest()
721
722
723

            mi = self._load_modelinfo_from_cache(module_hash)
            if mi is not None:
724
                logger.debug(
725
                    "Loaded model info for class %s.%s from cache",
726
727
728
                    self.module_name,
                    self.class_name,
                )
729
730
                return mi
            else:
731
                logger.debug(
732
                    "Cache model info for class %s.%s miss. Loading model instead.",
733
734
735
                    self.module_name,
                    self.class_name,
                )
736
737
738

        # Performed in another process to avoid initializing CUDA
        mi = _run_in_subprocess(
739
740
741
742
743
            lambda: _ModelInfo.from_model_cls(self.load_model_cls())
        )
        logger.debug(
            "Loaded model info for class %s.%s", self.module_name, self.class_name
        )
744
745

        # save cache file
746
747
        if module_hash is not None:
            self._save_modelinfo_to_cache(mi, module_hash)
748
749

        return mi
750

751
    def load_model_cls(self) -> type[nn.Module]:
752
753
754
755
756
757
758
759
        mod = importlib.import_module(self.module_name)
        return getattr(mod, self.class_name)


@lru_cache(maxsize=128)
def _try_load_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
760
) -> type[nn.Module] | None:
761
    from vllm.platforms import current_platform
762

763
    current_platform.verify_model_arch(model_arch)
764
765
766
    try:
        return model.load_model_cls()
    except Exception:
767
        logger.exception("Error in loading model architecture '%s'", model_arch)
768
        return None
769
770


771
772
773
774
@lru_cache(maxsize=128)
def _try_inspect_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
775
) -> _ModelInfo | None:
776
777
778
    try:
        return model.inspect_model_cls()
    except Exception:
779
        logger.exception("Error in inspecting model architecture '%s'", model_arch)
780
        return None
781
782


783
784
785
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
786
    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)
787

788
    def get_supported_archs(self) -> Set[str]:
789
        return self.models.keys()
790

791
792
793
    def register_model(
        self,
        model_arch: str,
794
        model_cls: type[nn.Module] | str,
795
    ) -> None:
796
797
798
        """
        Register an external model to be used in vLLM.

799
        `model_cls` can be either:
800

801
        - A [`torch.nn.Module`][] class directly referencing the model.
802
        - A string in the format `<module>:<class>` which can be used to
803
804
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
805
          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
806
        """
807
808
809
810
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

811
        if model_arch in self.models:
812
813
            logger.warning(
                "Model architecture %s is already registered, and will be "
814
815
816
817
                "overwritten by the new model class %s.",
                model_arch,
                model_cls,
            )
818
819
820
821
822
823

        if isinstance(model_cls, str):
            split_str = model_cls.split(":")
            if len(split_str) != 2:
                msg = "Expected a string in the format `<module>:<class>`"
                raise ValueError(msg)
824

825
            model = _LazyRegisteredModel(*split_str)
826
        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
827
            model = _RegisteredModel.from_model_cls(model_cls)
828
        else:
829
830
831
832
            msg = (
                "`model_cls` should be a string or PyTorch model class, "
                f"not a {type(model_arch)}"
            )
833
            raise TypeError(msg)
834

835
        self.models[model_arch] = model
836

837
    def _raise_for_unsupported(self, architectures: list[str]):
838
        all_supported_archs = self.get_supported_archs()
839

840
841
842
        if any(arch in all_supported_archs for arch in architectures):
            raise ValueError(
                f"Model architectures {architectures} failed "
843
844
                "to be inspected. Please check the logs for more details."
            )
845

846
847
848
849
850
851
852
853
        for arch in architectures:
            if arch in _PREVIOUSLY_SUPPORTED_MODELS:
                previous_version = _PREVIOUSLY_SUPPORTED_MODELS[arch]

                raise ValueError(
                    f"Model architecture {arch} was supported in vLLM until "
                    f"v{previous_version}, and is not supported anymore. "
                    "Please use an older version of vLLM if you want to "
854
855
                    "use this model architecture."
                )
856

857
858
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
859
860
            f"Supported architectures: {all_supported_archs}"
        )
861

862
    def _try_load_model_cls(self, model_arch: str) -> type[nn.Module] | None:
863
864
        if model_arch not in self.models:
            return None
865

866
        return _try_load_model_cls(model_arch, self.models[model_arch])
867

868
    def _try_inspect_model_cls(self, model_arch: str) -> _ModelInfo | None:
869
870
        if model_arch not in self.models:
            return None
871

872
873
874
875
876
877
        return _try_inspect_model_cls(model_arch, self.models[model_arch])

    def _try_resolve_transformers(
        self,
        architecture: str,
        model_config: ModelConfig,
878
    ) -> str | None:
879
880
881
        if architecture in _TRANSFORMERS_BACKEND_MODELS:
            return architecture

882
883
884
        auto_map: dict[str, str] = (
            getattr(model_config.hf_config, "auto_map", None) or dict()
        )
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900

        # Make sure that config class is always initialized before model class,
        # otherwise the model class won't be able to access the config class,
        # the expected auto_map should have correct order like:
        # "auto_map": {
        #     "AutoConfig": "<your-repo-name>--<config-name>",
        #     "AutoModel": "<your-repo-name>--<config-name>",
        #     "AutoModelFor<Task>": "<your-repo-name>--<config-name>",
        # },
        for prefix in ("AutoConfig", "AutoModel"):
            for name, module in auto_map.items():
                if name.startswith(prefix):
                    try_get_class_from_dynamic_module(
                        module,
                        model_config.model,
                        revision=model_config.revision,
901
                        trust_remote_code=model_config.trust_remote_code,
902
903
904
905
906
907
908
909
910
911
912
913
                        warn_on_fail=False,
                    )

        model_module = getattr(transformers, architecture, None)

        if model_module is None:
            for name, module in auto_map.items():
                if name.startswith("AutoModel"):
                    model_module = try_get_class_from_dynamic_module(
                        module,
                        model_config.model,
                        revision=model_config.revision,
914
                        trust_remote_code=model_config.trust_remote_code,
915
916
917
918
919
                        warn_on_fail=True,
                    )
                    if model_module is not None:
                        break
            else:
920
                if model_config.model_impl != "transformers":
921
922
923
924
925
926
927
                    return None

                raise ValueError(
                    f"Cannot find model module. {architecture!r} is not a "
                    "registered model in the Transformers library (only "
                    "relevant if the model is meant to be in Transformers) "
                    "and 'AutoModel' is not present in the model config's "
928
929
                    "'auto_map' (relevant if the model is custom)."
                )
930
931

        if not model_module.is_backend_compatible():
932
            if model_config.model_impl != "transformers":
933
                return None
934

935
936
            raise ValueError(
                f"The Transformers implementation of {architecture!r} "
937
938
                "is not compatible with vLLM."
            )
939

940
        return model_config._get_transformers_backend_cls()
941

942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
    def _normalize_arch(
        self,
        architecture: str,
        model_config: ModelConfig,
    ) -> str:
        if architecture in self.models:
            return architecture

        # This may be called in order to resolve runner_type and convert_type
        # in the first place, in which case we consider the default match
        match = try_match_architecture_defaults(
            architecture,
            runner_type=getattr(model_config, "runner_type", None),
            convert_type=getattr(model_config, "convert_type", None),
        )
        if match:
            suffix, _ = match

            # Get the name of the base model to convert
            for repl_suffix, _ in iter_architecture_defaults():
                base_arch = architecture.replace(suffix, repl_suffix)
                if base_arch in self.models:
                    return base_arch

        return architecture
967

968
969
    def inspect_model_cls(
        self,
970
        architectures: str | list[str],
971
        model_config: ModelConfig,
972
    ) -> tuple[_ModelInfo, str]:
973
974
        if isinstance(architectures, str):
            architectures = [architectures]
975
976
        if not architectures:
            raise ValueError("No model architectures are specified")
977
978

        # Require transformers impl
979
        if model_config.model_impl == "transformers":
980
            arch = self._try_resolve_transformers(architectures[0], model_config)
981
982
983
984
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)
985
        elif model_config.model_impl == "terratorch":
986
987
            model_info = self._try_inspect_model_cls("Terratorch")
            return (model_info, "Terratorch")
988

989
        # Fallback to transformers impl (after resolving convert_type)
990
991
992
993
994
995
        if (
            all(arch not in self.models for arch in architectures)
            and model_config.model_impl == "auto"
            and getattr(model_config, "convert_type", "none") == "none"
        ):
            arch = self._try_resolve_transformers(architectures[0], model_config)
996
997
998
999
1000
1001
1002
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

        for arch in architectures:
            normalized_arch = self._normalize_arch(arch, model_config)
1003
            model_info = self._try_inspect_model_cls(normalized_arch)
1004
            if model_info is not None:
1005
                return (model_info, arch)
1006

1007
        # Fallback to transformers impl (before resolving runner_type)
1008
1009
1010
1011
1012
        if (
            all(arch not in self.models for arch in architectures)
            and model_config.model_impl == "auto"
        ):
            arch = self._try_resolve_transformers(architectures[0], model_config)
1013
1014
1015
1016
1017
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

1018
        return self._raise_for_unsupported(architectures)
1019

1020
1021
    def resolve_model_cls(
        self,
1022
        architectures: str | list[str],
1023
        model_config: ModelConfig,
1024
    ) -> tuple[type[nn.Module], str]:
1025
1026
        if isinstance(architectures, str):
            architectures = [architectures]
1027
1028
        if not architectures:
            raise ValueError("No model architectures are specified")
1029
1030

        # Require transformers impl
1031
        if model_config.model_impl == "transformers":
1032
            arch = self._try_resolve_transformers(architectures[0], model_config)
1033
1034
1035
1036
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)
1037
        elif model_config.model_impl == "terratorch":
1038
1039
1040
1041
            arch = "Terratorch"
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
1042

1043
        # Fallback to transformers impl (after resolving convert_type)
1044
1045
1046
1047
1048
1049
        if (
            all(arch not in self.models for arch in architectures)
            and model_config.model_impl == "auto"
            and getattr(model_config, "convert_type", "none") == "none"
        ):
            arch = self._try_resolve_transformers(architectures[0], model_config)
1050
1051
1052
1053
1054
1055
1056
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

        for arch in architectures:
            normalized_arch = self._normalize_arch(arch, model_config)
1057
            model_cls = self._try_load_model_cls(normalized_arch)
1058
1059
            if model_cls is not None:
                return (model_cls, arch)
1060

1061
        # Fallback to transformers impl (before resolving runner_type)
1062
1063
1064
1065
1066
        if (
            all(arch not in self.models for arch in architectures)
            and model_config.model_impl == "auto"
        ):
            arch = self._try_resolve_transformers(architectures[0], model_config)
1067
1068
1069
1070
1071
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

1072
        return self._raise_for_unsupported(architectures)
1073

1074
1075
    def is_text_generation_model(
        self,
1076
        architectures: str | list[str],
1077
        model_config: ModelConfig,
1078
    ) -> bool:
1079
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1080
        return model_cls.is_text_generation_model
1081

1082
    def is_pooling_model(
1083
        self,
1084
        architectures: str | list[str],
1085
        model_config: ModelConfig,
1086
    ) -> bool:
1087
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1088
        return model_cls.is_pooling_model
1089

1090
1091
    def is_cross_encoder_model(
        self,
1092
        architectures: str | list[str],
1093
        model_config: ModelConfig,
1094
    ) -> bool:
1095
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1096
        return model_cls.supports_cross_encoding
1097

1098
1099
    def is_multimodal_model(
        self,
1100
        architectures: str | list[str],
1101
        model_config: ModelConfig,
1102
    ) -> bool:
1103
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1104
        return model_cls.supports_multimodal
1105

1106
    def is_multimodal_raw_input_only_model(
1107
        self,
1108
        architectures: str | list[str],
1109
        model_config: ModelConfig,
1110
    ) -> bool:
1111
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1112
        return model_cls.supports_multimodal_raw_input_only
1113

1114
1115
    def is_pp_supported_model(
        self,
1116
        architectures: str | list[str],
1117
        model_config: ModelConfig,
1118
    ) -> bool:
1119
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1120
        return model_cls.supports_pp
1121

1122
1123
    def model_has_inner_state(
        self,
1124
        architectures: str | list[str],
1125
        model_config: ModelConfig,
1126
    ) -> bool:
1127
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1128
        return model_cls.has_inner_state
1129

1130
1131
    def is_attention_free_model(
        self,
1132
        architectures: str | list[str],
1133
        model_config: ModelConfig,
1134
    ) -> bool:
1135
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1136
        return model_cls.is_attention_free
1137

1138
1139
    def is_hybrid_model(
        self,
1140
        architectures: str | list[str],
1141
        model_config: ModelConfig,
1142
    ) -> bool:
1143
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1144
1145
        return model_cls.is_hybrid

1146
1147
    def is_noops_model(
        self,
1148
        architectures: str | list[str],
1149
        model_config: ModelConfig,
1150
    ) -> bool:
1151
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1152
1153
        return model_cls.has_noops

1154
1155
    def is_transcription_model(
        self,
1156
        architectures: str | list[str],
1157
        model_config: ModelConfig,
1158
    ) -> bool:
1159
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1160
1161
        return model_cls.supports_transcription

1162
1163
    def is_transcription_only_model(
        self,
1164
        architectures: str | list[str],
1165
        model_config: ModelConfig,
1166
    ) -> bool:
1167
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1168
1169
        return model_cls.supports_transcription_only

1170

1171
1172
1173
1174
1175
1176
1177
1178
1179
ModelRegistry = _ModelRegistry(
    {
        model_arch: _LazyRegisteredModel(
            module_name=f"vllm.model_executor.models.{mod_relname}",
            class_name=cls_name,
        )
        for model_arch, (mod_relname, cls_name) in _VLLM_MODELS.items()
    }
)
1180
1181
1182
1183
1184

_T = TypeVar("_T")


def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
1185
1186
1187
1188
1189
    # NOTE: We use a temporary directory instead of a temporary file to avoid
    # issues like https://stackoverflow.com/questions/23212435/permission-denied-to-write-to-my-temporary-file
    with tempfile.TemporaryDirectory() as tempdir:
        output_filepath = os.path.join(tempdir, "registry_output.tmp")

1190
        # `cloudpickle` allows pickling lambda functions directly
1191
        import cloudpickle
1192

1193
        input_bytes = cloudpickle.dumps((fn, output_filepath))
1194
1195
1196

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
1197
1198
1199
        returned = subprocess.run(
            _SUBPROCESS_COMMAND, input=input_bytes, capture_output=True
        )
1200
1201
1202
1203
1204
1205

        # check if the subprocess is successful
        try:
            returned.check_returncode()
        except Exception as e:
            # wrap raised exception to provide more information
1206
1207
1208
            raise RuntimeError(
                f"Error raised in subprocess:\n{returned.stderr.decode()}"
            ) from e
1209

1210
        with open(output_filepath, "rb") as f:
1211
1212
1213
1214
1215
1216
            return pickle.load(f)


def _run() -> None:
    # Setup plugins
    from vllm.plugins import load_general_plugins
1217

1218
1219
1220
1221
1222
    load_general_plugins()

    fn, output_file = pickle.loads(sys.stdin.buffer.read())

    result = fn()
1223
1224
1225

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
1226
1227
1228


if __name__ == "__main__":
1229
    _run()